You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

MiniMax-H3 FL2VA + Larry v4 Turbo (merged transformer)

This repository contains the premerged FL2VA transformer only for native MiniMax-H3 first-frame, last-frame, and first+last-frame video-and-audio generation. It must be assembled with the non-transformer components from the pinned MiniMaxAI/MiniMax-H3 FL2VA partition before serving.

Pinned provenance

Base: MiniMaxAI/MiniMax-H3@42ed227ee7df40d41602854ae760620d6eb651fe, partition FL2VA
LoRA: larryvrh/MiniMax-H3-Turbo-Lora@43a74557ac3f6539db8e0f2a959d03feb7a81480
File: minimax_h3_turbo_v4_step600_ema.safetensors
LoRA SHA-256: 5f3a626cd72c93a8b9318d6760c510bc5092d2ab13aaba1f932c5bab07a416d3
Merge scale: 1.0
Merged targets: 259
Arithmetic: base dtype output with FP32 chunked W + B @ A accumulation

The Larry v4 model card recommends 4–8 denoiser evaluations. With SGLang's H3 scheduler contract, num_inference_steps: 9 produces 8 denoiser evaluations.

SGLang request mode

Serve the assembled model with --model-variant fl2va. A classic image-to-video request supplies one image condition with role: "keyframe" and frame_index: 0. Use -1 for a last-frame condition, or ordered [0, -1] conditions for both.

This derivative remains subject to the MiniMax-H3 community license linked above. The source LoRA is Apache-2.0 licensed. Repository ID: NaukNauk/minimax-h3-turbo-fl2va-merged.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for NaukNauk/minimax-h3-turbo-fl2va-merged

Base model

MiniMaxAI/MiniMax-H3

Adapter

(40)

this model